From: "Robert S. French" <rfrench@CS.STANFORD.EDU>
Date: Fri, 12 Aug 1994 10:18:20 -0700
Subject: Re: SUIF versus GCC
Message-Id: <9408121718.AA16555@Hawg.Stanford.EDU>
From: rjansen@voltaire.et.tudelft.nl (Roger Jansen)
Date: Fri, 12 Aug 1994 13:27:06 +0100 (METDST)
Reply-To: suif-talk@hawg
I would like to know how efficient the MIPS code is generated by scc. After
running both scc and gcc on severeral benchmarks I was a little disappointed
From the README file...
WHAT SUIF IS NOT
SUIF is not a production compiler. It is not suitable for replacing
your native compiler. ... SUIF will not generate scalar code that's
as good as your native compiler (unless you use the "C" backend and
actually compile with your native compiler...)
The research at Stanford these days is mainly on parallelism, data
distribution, or various other forms of high-level analysis. In
general we use the native C compiler as a backend because we are not
currently interested in scalar optimizations, instruction scheduling,
register allocation, etc. The MIPS code generator (mgen) and the
scalar optimizer (oynk) are from a previous version of SUIF (called
"old SUIF") and very little effort has been put into tuning them for
runtime performance.
There may be a new optimizer coming at some point. Various third
parties are working on code generators. But we at Stanford do not
have much time to work on parts of the compiler outside of our
research areas. This is where USERS come in. This is why SUIF is
easily expandable. We are relying on USERS to write any new passes
that they need, and then hopefully redistribute them to other users.
SUIF is still in the early stages of its public evolution. GCC has
been worked on by hundreds of people over quite a few years. It is
not at all surprising that GCC is tuned better, or looks more like a
production compiler. In fact, it would be surprising if that were NOT
the case.
So, please feel free to write a new optimizer or code generator for
us. We certainly won't complain :-)
Rob
From: tjiang@Synopsys.COM
Date: Fri, 12 Aug 94 10:13:36 -0700
Subject: Re: SUIF versus GCC
Message-Id: <199408121713.AA17617@gaea.synopsys.com>
Hi.
Since I was responsible for the optimizer in a previous life, I should
comment on this.
Scc's back-end (optimizer&code-gen) has not changed since 1991. It
has not kept up with the rest of SUIF and is completely out of tune.
There are no plans to tune the old back-end or add new optimizations
as any such effort would be better spent implementing the back-end
with new SUIF. I believe there are several ongoing efforts to build
new code generators using new SUIF.
Steve Tjiang
Synopsys Inc.
700 E Middlefield Road
Mountain View, CA 94043-4033
Phone: 415-694-4492
Fax: 415-965-8637
email: tjiang@synopsys.com
Date: Fri, 12 Aug 94 13:27:06 BST
Subject: SUIF versus GCC
From: rjansen@voltaire.et.tudelft.nl (Roger Jansen)
X-Mailer: ELM [version 2.4 PL21]
Content-Type: text
Content-Length: 739
Sender: owner-suif-talk%hawg@hawg.Stanford.EDU
Precedence: bulk
Reply-To: suif-talk%hawg@hawg.Stanford.EDU
I would like to know how efficient the MIPS code is generated by scc. After
running both scc and gcc on severeral benchmarks I was a little disappointe
d
by the results. I have used both compilers with the -O2 option.
Here are some results I have found with the tool 'pixie' on a DEC-station:
scc gcc
# instr. # cycles # instr. # cycles
bubble.c 2,537,566 2,557,642 1,413,151 1,429,727
dhrystone.c 7,171,379 7,172,164 5,021,181 5,121,965
quickrand.c 1,196,773 1,392,029 1,083,772 1,244,028
saxpy.c 2,300,318 2,300,318 1,400,356 1,400,356
stanford.c 28,249,625 29,725,447 21,083,713 22,559,337
Roger Jansen.
From: rjansen@voltaire.et.tudelft.nl (Roger Jansen)
Date: Fri, 12 Aug 1994 13:27:06 +0100 (METDST)
Subject: SUIF versus GCC
Message-Id: <199408121127.AA110880826@voltaire.et.tudelft.nl>
I would like to know how efficient the MIPS code is generated by scc. After
running both scc and gcc on severeral benchmarks I was a little disappointed
by the results. I have used both compilers with the -O2 option.
Here are some results I have found with the tool 'pixie' on a DEC-station:
scc gcc
# instr. # cycles # instr. # cycles
bubble.c 2,537,566 2,557,642 1,413,151 1,429,727
dhrystone.c 7,171,379 7,172,164 5,021,181 5,121,965
quickrand.c 1,196,773 1,392,029 1,083,772 1,244,028
saxpy.c 2,300,318 2,300,318 1,400,356 1,400,356
stanford.c 28,249,625 29,725,447 21,083,713 22,559,337
Roger Jansen.